/GFANC-Bayes

Delayless Generative Fixed-filter Active Noise Control based on Deep Learning and Bayesian Filter

Primary LanguageJupyter Notebook

Delayless Generative Fixed-filter Active Noise Control based on Deep Learning and Bayesian Filter (GFANC-Bayes)

This is the code of the paper 'Delayless Generative Fixed-filter Active Noise Control based on Deep Learning and Bayesian Filter' accepted by IEEE Transactions on Audio, Speech and Language Processing (TASLP). You can find the paper at Researchgate or Ieee xplore.

HIGHLIGHTS:

  1. To address this limitation of the SFANC method and generate more appropriate control filters, a generative fixed-filter active noise control approach based on Bayesian filter (GFANC-Bayes) is proposed in this paper.

  2. The GFANC-Bayes method can automatically generate suitable control filters by combining sub control filters. The combination weights of sub control filters are predicted via a 1D CNN. The predicted combination weights are then filtered by a Bayesian filtering module, which exploits the correlation information between adjacent noise frames to improve the prediction accuracy and robustness.

  3. To achieve delayless noise control, the co-processor operates at the frame rate while the real-time controller performs at the sample rate in parallel.

How to use the code:

If you don't want to retrain the 1D CNN ('M5_Network.py'), the trained model can be found in 'models/M6_res_Synthetic.pth', you can easily run the 'Main-GFANC-Bayes.ipynb' file to get the noise reduction results.

The 1D CNN is trained using a synthetic noise dataset, its label files are 'Soft_Index.csv' and 'Hard_Index.csv'. The entire dataset is available at https://drive.google.com/file/d/1hs7_eHITxL16HeugjQoqYFTs-Cm7J-Tq/view?usp=sharing

Especially, the pre-trained sub control filters are obtained on synthetic acoustic paths, where the primary and secondary paths are bandpass filters. If you want to use the GFANC-Bayes system on new acoustic paths only requires obtaining the corresponding broadband control filter and decomposing it into sub control filters. Noticeably, the trained 1D CNN in the GFANC-Bayes system remains unchanged. The detailed information can be found in Section 'Noise Cancellation on Measured Acoustic Paths' in the paper.

RELATED PAPERS: